CAM-based search engine devices having advanced search and learn instruction handling
CAM-based search engine devices operate to reduce the occurrence of duplicate learned entries within a CAM database when processing search and learn (SNL) instructions. A search engine device may be configured to support processing of first and second immediately consecutive and equivalent SNL instructions as a first SNL instruction and a second search and search instruction, respectively. This processing is performed in order to block an addition of a duplicate learned entry within a database in the search engine device. The search engine device may also be configured to selectively block processing of the second SNL instruction as a second search and search instruction in response to detecting the database as full when the first SNL instruction is processed.
Latest Integrated Device Technology, Inc. Patents:
- Wireless power Tx to Tx conflict and protection
- Wireless power receiver configurable for LDO or buck operation
- Light detection and ranging system and method for operating a light detection and ranging system
- Sliding window and DC offset correction technique for pulse doppler radar systems
- Inductive position sensor for electronic throttle control
This application is a continuation-in-part (CIP) of U.S. application Ser. No. 10/714,680, filed Nov. 14, 2003 now U.S. Pat. No. 7,120,733, which is a continuation-in-part of U.S. application Ser. No. 10/698,246, filed Oct. 31, 2003 now U.S. Pat. No. 7,082,493, which claims priority to U.S. Provisional Application Ser. No. 60/516,178, filed Oct. 31, 2003, the disclosures of which are hereby incorporated herein by reference.
FIELD OF THE INVENTIONThe present invention relates to integrated circuit devices that support search operations and, more particularly, to CAM-based search engine devices and methods of operating same.
BACKGROUND OF THE INVENTIONConventional network processor units (NPU) may be interfaced to integrated IP coprocessors (IIPC) in a manner that enables both SRAMs and IIPCs to be operated on the same memory mapped bus. As illustrated by
The IIPC 30 is illustrated as including a content addressable memory (CAM) core 36 and logic 38 that couples the CAM core 36 to the memory mapped interface. This memory mapped interface is illustrated as including read control logic 32 and write control logic 34. The write control logic 34 is configured to receive an address ADDR[21:0], a write enable signal WE_N[1:0], input data DATAIN[15:0] and input parameters PARIN[1:0]. The read control logic 32 is configured to receive the address ADDR[21:0] and a read enable signal RE_N[1:0] and generate output data DATAOUT[15:0] and output parameters PAROUT[1:0]. Like the SRAM controller within the NPU 10, this memory mapped interface is based on FIFO communication. The IIPC 30 performs operations using the input data DATAIN[15:0] and input parameters PARIN[1:0] and then passes back result values to the NPU 10. The timing between the receipt of the input parameters and the return of the corresponding result values is not fixed. Instead, it is determined by the amount of time the IIPC 30 requires to execute the specified instruction and depends on the number and type of other instructions currently pending within the IIPC 30.
These pending instructions are initially logged into respective instruction control registers 50 that support a plurality of separate contexts (shown as a maximum of 128). These instructions may be processed in a pipelined manner. The result values generated at the completion of each context are provided to respective result mailboxes 40. The validity of the result values within the mailboxes 40 is identified by the status of the done bit within each result mailbox 40. Accordingly, if a read operation is performed before the result values are ready, the NPU 10 will be able to check the validity of the done bit associated with each set of result values to determine whether the corresponding values of valid. However, because there can be multiple contexts in progress within the IIPC 30 at any given time and because the completion of the contexts does not necessarily occur in the same sequence as the requests were made, the NPU 10 may need to regularly poll the result mailboxes 40 at relatively high frequency to obtain new results as they become valid. Unfortunately, such regular polling can consume a substantial amount of the bandwidth of instructions that are issued to the IIPC 30 and lead to relatively high levels of operational inefficiency when the IIPC 30 is running a large number of contexts. Thus, notwithstanding the IIPC 30 of
Referring now to
The internal CAM core 330 is illustrated as a ternary CAM core that contains a data array and a mask array 328. This CAM core 330 may be configurable into a plurality of independently searchable databases. General and database configuration registers 318 are also provided along with global mask registers GMRs 320. These registers provide data to instruction loading and execution logic 332, which may operate as a finite state machine (FSM). The instruction loading and execution logic 332 communicates with the CAM core 330 and the result logic 334. If the IIPC 300 is configured to support a depth-cascaded mode of operation, a cascade interface 338 may be provided for passing data and results to (and from) another IIPC (not shown). The instruction loading and execution logic 332 may also pass data to and from an external memory device, via an SRAM interface 336.
The aging logic 321 is illustrated as including two memory arrays: an age enable array 322 and an age activity array 324. These memory arrays may have bit positions that map directly to entries within the CAM core 330. Thus, if the CAM core 330 has 128 k entries (e.g., x72 entries), then the age enable array 322 and age activity array 324 may each have a capacity of 128 k bits. The illustrated aging logic 321 may operate with the instruction loading and execution logic 332 to (i) reset age activity bits that have been previously set within the age activity array 324 in response to successful search operations and (ii) age out entries associated with previously reset activity bits by invalidating the corresponding entries.
The aging operations may include periodically inserting an aging instruction into an instruction pipeline within the IIPC 300. As illustrated by the global and database aging request circuit 350 of
Conventional CAM-based search engine devices may also be configured to perform search and learn (SNL) operations that include an initial search operation and, if necessary, a subsequent learn operation of a new entry into a CAM core. As illustrated by FIG. 5 of U.S. Pat. No. 6,219,748, operations may be performed to decode a learn instruction and load comparand data (e.g., search key) into a comparand register within a search engine device. An operation is then performed to search a CAM array using the comparand data as a search word. If a match occurs in response to the search operation, then operations associated with generating conventional search results are performed. However, if no match is present and the CAM array is not full, then the comparand may be written as a CAM entry into an internally generated next free address within the CAM array. This write operation into a next free address is treated as a learn operation as opposed to a conventional write operation that includes an externally supplied write address. The next free address associated with the newly written CAM entry is then reported to a command host. Related CAM-based search engine devices that describe learn operations are also disclosed in U.S. Pat. Nos. 6,148,364 and 6,240,485.
SUMMARY OF THE INVENTIONCAM-based search engine devices operate to reduce the occurrence of duplicate learned entries within a CAM database when processing search and learn (SNL) instructions. According to first embodiments of the present invention, a search engine device is configured to support processing of equivalent SNL instructions in a manner that reduces the occurrence of duplicate learned entries within a CAM database even when the equivalent SNL instructions are received by the search engine device during immediately consecutive time intervals. SNL instructions are treated as equivalent when they are associated with the same search key and directed at the same database(s) within a CAM core. In particular, a search engine device is configured to process first and second immediately consecutive and equivalent SNL instructions as: (i) a first SNL instruction, and (ii) a second search and search (SNS) instruction, respectively. This preferred processing is performed in order to block an addition of a duplicate learned entry within a database in the search engine device. The search engine device may also be configured to support processing of the first and second immediately consecutive and equivalent SNL instructions in a manner that prevents a failure to learn a new entry when the corresponding database is full. In particular, the internal conversion of an SNL instruction into an SNS instruction can be suspended when the corresponding database is full.
According to second embodiments of the present invention, a CAM-based search engine device is provided that is configured to convert a learn portion of a search-and-learn (SNL) instruction associated with a search key into a search operation using the search key. This conversion operation is performed in response to detecting a prior equivalent learn of the search key in the search engine device. The search engine device may also include a SNL cache memory device that is configured to store search keys that accompany SNL instructions and marker information (e.g., flags) that identifies whether the stored search keys are duplicates of other search keys within the SNL cache memory device. This SNL cache memory device may operate as a first-in first-out (FIFO) memory device.
Still further embodiments of the present invention include methods of operating CAM-based search engine devices. These methods include searching a database within a CAM core with a search key to detect the presence of a hit or miss condition. Then, if a miss condition has been detected, as step is performed to check the search key to determine whether is has been marked as a duplicate. If the search key has been marked as a duplicate, then the checking step is followed by the step of converting a learn portion of the SNL instruction into a search operation that results in another search of the database with the search key.
The present invention now will be described more fully herein with reference to the accompanying drawings, in which preferred embodiments of the invention are shown. This invention may, however, be embodied in many different forms and should not be construed as being limited to the embodiments set forth herein; rather, these embodiments are provided so that this disclosure will be thorough and complete, and will fully convey the scope of the invention to those skilled in the art. Like reference numerals refer to like elements throughout and signal lines and signals thereon may be referred to by the same reference characters. Signals may also be synchronized and/or undergo minor boolean operations (e.g., inversion) without being considered different signals. Moreover, when a device or element is stated as being responsive to a signal(s), it may be directly responsive to the signal(s) or indirectly responsive to the signal(s) (e.g., responsive to another signal(s) that is derived from the signal(s)).
Referring now to
The result status select register(s) 70 is a 128-bit programmable register that generates a result status select signal RSS<0:127>. This signal operates to select one of two indication circuits for receipt of active bits within the result status signal RS<0:127>. These indication circuits are illustrated as an interrupt indication circuit 60a and a non-interrupt indication circuit 60b. The interrupt indication circuit 60a includes an interrupt generator 64 that generates an interrupt INT to the command host 140 via the memory mapped interface 130. The interrupt generator 64 may also generate interrupts in response to other activity within the control circuit, according to a predefined protocol. In contrast, the non-interrupt indication circuit 60b generates an asynchronous aggregate result status signal (ARS) to the command host 140 via the memory mapped interface 130. This ARS signal is configured to have a leading edge that occurs when a first one of a selected plurality of contexts is completed and an active level that is held so long as at least one of the selected plurality of contexts remains completed (i.e., done status bit remains set).
The interrupt indication circuit 60a has a first bank 62a of AND gates that output to an OR gate 68a. The non-interrupt indication circuit 60b has a second bank 62b of AND gates that output to an OR gate 68b. When one or more bits of the result status select signal RSS<0:127> are set high to logic 1 levels, then the corresponding result status signals RS<0:127> are passed to the inputs of the OR gate 68a. If any of these result status signals are switched to active logic 1 values, then the output of the OR gate 68a will switch and cause the interrupt generator 64 to produce an interrupt INT at the memory mapped interface 130. But, when one or more bits of the result status select signal RSS<0:127> are set low to logic 0 levels, then the corresponding result status signals RS<0:127> are passed to the input of the OR gate 68b. Accordingly, if the result status select signal RSS<0:127> is set so that RSS<0:127>=<00000 . . . 0000>, then the aggregate result status signal at the output of the OR gate 68b will be switched high (or held high) whenever any of the result status bits RS<0:127> is set high to indicate the completed state of a respective context. Alternatively, if the result status select signal RSS<0:127> is set so that RSS<0:127>=<11111 . . . 1111>, then the signal at the output of the OR gate 68a will be switched high (or held high) whenever any of the result status bits RS<0:127> is set high to indicate the completed state of a respective context. In this manner, the result status select register 70 provides programmable control over how the result status signals are to be reported to the command host 140.
Based on the above-described configuration of the control circuit, the completion of any context within the IIPC 100 will result in the transfer of result values from the scheduler, state machine and logic 110 to a corresponding result mailbox 90. Assuming this context represents a first-to-finish operation (e.g., lookup within the CAM core), then the setting of the respective done bit within the result mailbox 90 will result in the latching of this done information by the result status register(s) 80. If this done information relates to context 0, then the result status signal RS<0:127> will equal <10000 . . . 000>. If the result status select register is set so that the result status select signal RSS<0:127> equals <0XXXXX . . . X>, where X represents a “don't care” for purposes of this example, then the aggregate result status signal ARS will be set to an active high level and passed from the memory mapped interface 130 to the command host 140. Alternatively, if the result status select register is set so that the result status select signal RSS<0:127> equals <1XXXXXX . . . X>, then the output of the OR gate 68a within the interrupt indication circuit 60a will switch high. This active high signal at an input of the interrupt generator 64 will result in the generation of an interrupt that passes to the memory mapped interface 130 and the command host 140.
In response to the generation of an interrupt INT or an active high aggregate result status signal ARS, the command host 140 may issue an operation to read the result status register 80. This operation includes generating an address ADDR[23:0] to the memory mapped interface 130. The fields of this address are illustrated by TABLE 1. The two most significant bits of the address operate to select the particular IIPC 100 for which the read operation is destined. The seven address bits ADDR<21:15> identify a particular context within a range of 128 possible contexts. The eleven address bits ADDR<4:14> are not used. The address bit ADDR<3> represents a result status identifier (RES_STATUS). If this bit is set to a first logic value (e.g., 0), then an entry within the result mailbox 90 associated with the designated context is to be read back to the command host 140. On the other hand, if the result status identifier is set to a second logic value (e.g., 1), then a designated portion of the result status register 80, which identifies the value of 32 result status signals, is to be read back to the command host. The final 3-bit portion of the address, shown as ADDR<2:0>, identifies an entry value. As illustrated by TABLE 2, this entry value identifies one of eight entries to be read from the designated result mailbox 90 when the result status identifier RES_STATUS is set to a logic 0 value. Alternatively, the entry value identifies one of four portions of the result status register 80 to read from when the result status identifier is set to a logic 1 value. In this manner, four consecutive read operations may be performed to enable the command host to read the entire contents of the result status register 80 and thereby readily identify which ones of the 128 result mailboxes 90 contain valid result values.
Referring now to
Referring now to
A clock generator circuit 530 and reset logic circuit 532 are also provided. The clock generator circuit 530 may include a delay and/or phase locked loop circuit that is configured to generate internal clock signals that are synchronized with an external clock signal EXTCLK. The reset logic circuit 532 may be configured to perform reset operations when the device 500 is initially powered up or after a chip reset event has occurred. An SRAM interface 534 may also be provided to enable transfer of data to and from an external memory device (e.g., associated SRAM). A cascade interface 536 is provided to support depth-cascading between the search engine device 500, operating as a “master” device, and a plurality of additional “slave” search engine devices that may be coupled together as illustrated and described more fully hereinbelow with respect to
First and second context sensitive logic circuits 510 and 512 are coupled to the first and second memory mapped interfaces 504 and 506, respectively. These context sensitive logic circuits 510 and 512 are illustrated as including instruction FIFOs and results mailboxes. The context sensitive logic circuits 510 and 512 may also includes results status circuits that are configured to generate respective aggregate result status signals (ARS) and interrupts, as described more fully hereinabove with respect to
An instruction loading and execution logic circuit 524 is provided with an instruction scheduler 527 and a search and learn (SNL) cache 525. This logic circuit 524 may perform the functions of a finite state machine (FSM) that controls access to the CAM core 522 and utilizes resources provided by specialized function registers 514, global mask registers 516, parity generation and checking circuitry 520 and an aging control logic circuit 518. The SNL cache 525 may support the performance of search and learn operations within the CAM core 522. During search operations, the instruction loading and execution logic circuit 524 provides the CAM core 522 with search words that may be derived from search keys received at a memory mapped interface. In response to a search operation, the CAM core 522 may generate a plurality of hit signals that are encoded to identify an address of a highest priority matching entry within the CAM core 522. This address may also be encoded as an absolute index that specifies the location of the highest priority matching entry with a multi-chip search machine. In some embodiments, the address may be provided to an index translation logic circuit 526 (ITL). This index translation logic circuit 526 may modify the addresses relative to a selected database to thereby create database relative indexes. Alternatively, the addresses may be modified relative to an NPU-attached associated SRAM to thereby create memory pointer indexes. A results logic circuit 528 is also provided. The results logic circuit 528 is configured to pass results values from the index translation logic circuit 526, the instruction loading and execution logic circuit 524 and the cascade interface 536 to results mailboxes associated with the context sensitive logic circuits 510 and 512 and the interface logic circuit 508.
The aging control logic circuit 518 is illustrated as including a plurality of memory devices, which may be updated as each entry is written into the CAM core 522 and during periodic aging operations. These memory devices include a quad arrangement of SRAM memory arrays 700a–700d, as illustrated more fully by
The data within the age enable memory array 700a identifies which CAM core entries are subject to aging. For example, each bit position within the age enable memory array 700a that is set to a logic 1 value (or logic 0 value) may reflect a corresponding CAM core entry that is subject to (or not subject to) aging. Each bit position within the age activity memory array 700b may reflect whether a corresponding CAM core entry has remained active since the time it was first written into the CAM core 522. For example, a logic value of 1 may reflect an active CAM core entry that has been the subject of a “hit” during a search operation (or one that has been relatively recently written to the CAM core) and a logic value of 0 may reflect an inactive CAM core entry that is ready to be aged out of the CAM core 522. Some of the automated aging operations associated with the age enable and age activity memory arrays 700a–700b are described more fully hereinabove with reference to
The age report enable memory array 700c reflects which entries are to be reported to a command host in response to being aged out of the CAM core 522. In the event a report only aging feature is provided on a global (i.e., full CAM core), per database and/or per entry basis, the age report enable memory array 700c may also identify those entries that have exceeded an activity-based aging threshold but have not undergone a final aging out operation (i.e., their valid bits have not been reset to an invalid condition). Thus, a bit position having a logic value of 1 within the age report enable memory array 700c may identify a corresponding CAM core entry as being subject to age reporting. In contrast, a bit position having a logic value of 0 within the age report enable memory array 700c may identify a corresponding CAM core entry as not being subject to age reporting when the entry is aged out of the CAM core 522.
The age FIFO select memory array 700d reflects where an entry, which is already the subject of age reporting, is reported to upon being aged out of the CAM core 522. By using one bit per CAM entry, one of two possible age reporting locations is possible. These two age reporting locations include a first FIFO (FIFO 0) and a second FIFO (FIFO 1), which are located within the aging control logic circuit 518. These FIFOs may each have a capacity of 255 entries. By using a larger memory array, which supports two or more bits per CAM entry, a greater number of age reporting locations may be identified by the age FIFO select memory array 700d. These first and second FIFOs may be accessed from any of the illustrated interfaces.
The instruction loading and execution logic circuit 524 also operates to control the periodic reporting of the addresses/indexes of the entries from the reporting locations (i.e., FIFO 0 and FIFO 1) to a command host. The phrase “periodic reporting” includes regularly spaced or intermittent reporting that is initiated by the command host or possibly initiated by the IIPC. These reporting operations are performed with the assistance of a plurality of the specialized function registers 514. These registers 514 include a first level count register and a second level count register. The first level count register is configured to maintain a count of unreported addresses that are stored in aging FIFO 0 and the second level count register is configured to maintain a count of unreported addresses that are stored in aging FIFO 1. The registers 514 also includes a first level configuration register and a second level configuration register. The first level configuration register is configured to maintain a programmable threshold count value that specifies how many addresses can be stored in aging FIFO 0 before the control circuit issues an interrupt to the command host (e.g., NPU 0) to thereby prompt the command host to issue a read request for the addresses stored within aging FIFO 0. Similarly, the second level configuration register is configured to maintain a programmable threshold count value that specifies how many addresses can be stored in aging FIFO 1 before the control circuit issues an interrupt to the command host (e.g., NPU 1) to thereby prompt the command host to issue a read request for the addresses stored within aging FIFO 1. The registers 514 may also include a first interrupt timer register that operates as a timer to support generation of an interrupt to the command host when no new addresses have been reported to aging FIFO 0 during a programmed time interval and at least one unreported address is stored within aging FIFO 0. This first interrupt timer is used so that the command host (e.g., NPU 0) is aware of the presence of at least one address within aging FIFO 0, even though the threshold count value stored in the first level configuration register has not been exceeded. A second interrupt timer register is also provided to operate in a similar manner with respect to aging FIFO 1.
Aging operations performed by the control circuit of
Blocks 610–616 illustrate a sequence of operations that may be performed to generate each aging operation request on a global basis within the search engine device. At Block 610, a countdown operation is commenced in a global aging register and a check is continuously made at Block 612 to determine whether a countdown operation has completed. If so, an aging operation is requested (see, Block 602) and the global aging register count is reloaded into the global aging register, Block 616.
Blocks 618–624 illustrate operations that may be used to generate age service requests for respective databases. If a CAM core is configured to support a maximum of 16 databases, then sixteen sets of operations corresponding to Blocks 618–624 are performed in parallel at potentially different frequencies. As illustrated by Block 618, a countdown operation is performed on a database aging register at a specified frequency. When the count reaches zero, an age service request is issued for the corresponding database Blocks 620–622. At Block 624, the corresponding database aging register count is reinitialized and the operations are repeated. The database aging register count values should be sufficiently high to prevent a backlog of age service requests for a given database when the round-robin arbitration of the database age servicing requests is performed, Block 606.
As illustrated by
If report-only aging is not enabled, then the selected entry is removed from its database (e.g., the entry is marked as invalid using a CLEAR VALID instruction that causes an access to the CAM core 522), Block 638. An entry may be marked as invalid by resetting the validity bit for the entry. Alternatively, a predetermined data string having a validity bit that is set to an invalid state may be written over the aged out entry. This may be particularly helpful in those embodiments that support background error detection and/or correction with parity and/or Hamming code bits. In some cases, the value of the validity bit may influence the value of the parity and/or Hamming code bits and merely resetting the validity bit when performing an age out operation may cause the entry to be improperly detected as invalid (and then possibly corrected by setting the validity bit to a valid state) during a background error detection and/or correction operation. To prevent the unintentional correction of an aged out entry, the predetermined data string having correct parity and/or Hamming code bits may be used as a default word that is to be written over every entry that is to be aged out of the CAM core.
As illustrated by Block 639, the corresponding age enable bit within the age enable memory array 700a is cleared so that the selected entry is no longer evaluated for aging (see, Block 632). A check is then made to determine whether the selected entry is subject to reporting to the command host (e.g., NPU 0, NPU 1 or PCI), Block 640. This check can be performed by evaluating the corresponding bit position within the age report enable memory array 700c. Accordingly, even if a selected entry is identified at Block 637 as being subject to report-only aging at a global or per database level, the check at Block 640 may override these settings for a given entry.
If the aged entry is subject to reporting, then the age reporting enable setting for the entry is cleared, Block 641, and the address/index of the entry is added (i.e., “reported”) to either FIFO 0 or FIFO 1, Block 642. The destination FIFO to which the aged entry is added is controlled by the value of the corresponding bit position within the age FIFO select memory array 700d. If the aged entry is reported to FIFO 0, then the identity of the aged out entry will ultimately be read from one of the memory mapped interfaces. Alternatively, if the aged entry is reported to FIFO 1, then the identity of the aged entry will ultimately be read from another one of the memory mapped interfaces. The timing of these read operations is a function of the timing of when the respective command hosts (e.g., NPU 0, NPU 1 or PCI), which issue the FIFO read instructions, receive corresponding interrupts that identify FIFO 0 or FIFO 1 as being sufficiently full. In the event FIFO 0 or FIFO 1 becomes completely full before being emptied by a command host, the instruction loading and execution logic 524 may operate to suspend age reporting or even operate to suspend all aging operations until such time as the age reporting FIFOs have been emptied.
The control circuit within the search engine device 500 may also be configured to fill FIFO 0 and FIFO 1 with the addresses of entries that have been aged out of other search engine devices. For example, when the illustrated search engine device 500 is configured as a master search engine device within a depth-cascaded search machine, the cascade interface 536 will operate to pass the indexes of aged out entries from one or more “slave” search engine devices to the aging FIFOs within the master search engine device. Accordingly, as illustrated by
Referring now to
The logic circuit 524 is illustrated as receiving a plurality of instructions. According to one environmental example, these instructions may include a search instruction (with Search Key 0) from IF2, a write instruction (with Search Key 1) from IF1, and two equivalent SNL instructions (with Search Key 2) from IF0 that are pipelined into the search engine device 900 in consecutive sequence. In alternative examples, these two equivalent SNL instructions may be received from different instruction FIFOs and be associated with different contexts. The logic circuit 524 arbitrates to determine the sequence of handling the competing instructions and access to the CAM core 522. As described herein, SNL instructions are deemed equivalent when they are associated with same search keys and directed at the same database(s) within the CAM core 522.
The handling of the two equivalent SNL instructions by the logic circuit 524 and CAM core 522 of
As illustrated by
The operation to add a new search key to the SNL cache memory device 525 may constitute a “push” operation onto a FIFO memory “stack.” An operation is then performed to determine whether the newly added search key is a duplicate of a search key currently residing in the SNL cache memory device 525, Block 1012. If a duplicate search key is not present, then the search key is marked with a learn instruction, Blocks 1014 and 1016. However, if a duplicate search key is present, then the search key is marked with a search instruction instead of a learn instruction, Blocks 1014 and 1018. These marking operations may cause the generation of opposite flag values associated with each entry in the FIFO memory device (e.g., flag=1 means the search key is marked with a search instruction and flag=0 means the search key is marked with a learn instruction). These flag values may constitute “marker” information.
Returning to Block 1020, if the search portion of the SNL instruction results in a hit condition, then this hit condition and a corresponding index of a matching entry are returned to a results logic circuit (see, Block 528 in
The operations illustrated by
In
The first SNL instruction SNL_1 and search key (Search Key 2) are transferred to the CAM core 522 and the search key is transferred to the SNL cache 525, Blocks 1008 and 1010. A search of the SNL cache 525 is then performed to detect the presence of a duplicate search key. This search of the cache results in a miss, Block 1014. As illustrated by Block 1016, the search key is marked with a learn instruction, which means a flag may be set that designates the search key as one that is to accompany a learn instruction when it is subsequently read from the SNL cache 525. At Block 1020, a check is made to determine whether a search of the CAM core 522 resulted in a hit or miss. Because the CAM core 522 did not contain the search key (i.e., Search Key 2), the check will result in a miss result. Then, at Block 1024, the flag associated with the search key in the SNL cache 525 will be checked to see whether it designates an attached learn instruction (key is not marked as a duplicate) or whether it designates an attached search instruction (key is marked as a duplicate). Because the search key is marked with a learn instruction, the search key and learn instruction are transferred to the CAM core 522 and the search key (Search Key 2) is learned, Block 1028. Thus, the first SNL instruction results in a search operation followed by a learn operation. In response, the CAM core 522 is updated with a new entry (Search Key 2).
At possibly the same time as the first search operation of SNL_1 is being checked at Block 1020, the second SNL instruction SNL_2 and search key (Search Key 2) are transferred to the CAM core 522 and the search key is transferred to the SNL cache 525, Blocks 1008 and 1010. At Blocks 1012 and 1014, the search key will be marked as a duplicate because the earlier equivalent search key is still held by the SNL cache 525. This means a flag may be set that designates the search key as one that is to accompany a search instruction when it is subsequently read from the SNL cache 525, Block 1018.
At Block 1020, a check is made to determine whether a search of the CAM core 522 resulted in a hit or miss. Because the CAM core 522 has not yet learned the search key as a result of the SNL_1 instruction, this check will result in another miss result. Then, at Block 1024, the flag associated with the search key in the SNL cache 525 will be checked to see whether it designates an attached learn instruction (key is not marked as a duplicate) or whether it designates an attached search instruction (key is marked as a duplicate). Because the search key is marked with a search instruction, the duplicate search key and search instruction are transferred to the CAM core 522 and the search operation is performed, Block 1026. At Block 1030, the results of this second search operation associated with SNL_2 are processed. These results include an indication of a hit condition (because of the earlier learn operation associated with SNL_1) and an index of the matching entry. Accordingly, rather than having two SNL instructions result in duplicate learning events into a database (because they arrive too close in time for the first SNL instruction to take effect before the search portion of the second SNL instruction is performed), the second SNL instruction is converted into a search and search (SNS) instruction, which results in a hit condition and returns an address of the learned entry back to a results mailbox.
This sequence of operations is also illustrated by
At
In
One potential limitation associated with the above-identified operations has to do with the processing of equivalent SNL instructions when a corresponding database to which the SNL instructions apply is full. In such a case, the first SNL instruction will not result in a successful learn operation and the marking of duplicate entries within the SNL cache 525 may result in repeated searches of the CAM core 522 and possibly an absence of learn instructions to update the CAM core 522 when the corresponding database is finally free to accept a new entry. To avoid this potential limitation, operations may be performed to clear one or more duplicate flag settings associated with the related SNL cache entries when a corresponding database (to which the search key is to be learned) is full. In particular, a configuration register associated with the registers 514 (see,
Referring now to
However, if the incoming instruction is a LEARN instruction, then a search is made of the I_CACHE to detect the presence of an equivalent search key (i.e., same key value and same database identifier), Block 1206b. At Block 1208b, a check is made to determine whether an equivalent search key was detected based on the search at Block 1206b. If an equivalent search key is not present, then the search key is added as an entry to the I_CACHE and a duplicate bit associated with the search key entry is set (e.g., duplicate bit is set to 1 binary), Block 1214. The instruction insertion operations starting at Block 1224 are then performed. But, if an equivalent search key is present based on the check at Block 1208b, then a check is made of the I_CACHE to determine whether a duplicate bit for the search key has been asserted, Block 1212. If not, then the duplicate bit is set (i.e., asserted) at Block 1216 and control is passed to Block 1224. If yes, the LEARN instruction is blocked, Block 1222, and control is passed to Block 1224, where the CAM core may experience of no-op cycle. Although the learn instruction is blocked, additional operations may be performed to update a results mailbox to indicate that the search key associated with the blocked instruction was previously learned.
Referring again to Block 1202, if a SEARCH instruction is detected, then control is passed to Block 1206a, where a search of the I_CACHE is performed to detect an equivalent search key. If an equivalent search key is not present, Block 1208a, then control is passed to Block 1224. But, if an equivalent search key is present, then a check is made to determine whether the corresponding duplicate bit is asserted, Block 1210. If a duplicate bit is asserted, then control is passed to Block 1224. If the duplicate bit is not asserted, then the duplicate bit is set, Block 1218, and the SEARCH instruction is converted into a LEARN instruction, Block 1220, before control is passed to Block 1224.
Once an instruction has been inserted into the instruction pipeline at Block 1224, the instruction (e.g., SEARCH, LEARN, WRITE, READ, etc.) is performed within the CAM core, Block 1226. If the result of a CAM core operation indicates that a search has been performed and a MISS result has been generated, Block 1228, then the corresponding search key is added to the I_CACHE, Block 1230, and control is passed to Block 1232 where results of a CAM core access are processed. (See, e.g.,
The operations illustrated by
The second LEARN instruction also passes to Block 1208b where its search key is compared with the entries in the I_CACHE. Because of the earlier I_CACHE update caused by the first LEARN instruction, the check at Block 1208b results in an affirmative answer. A check to determine whether the corresponding duplicate bit has been asserted is then performed, Block 1212. This check also results in an affirmative answer (based on the earlier learn of the equivalent search key) and control is passed to Block 1222. At Block 1222, the second LEARN instruction is blocked in order to prevent a a duplicate learn event from occurring within the CAM core.
In a second example, two equivalent SEARCH instructions, which have the same key and are directed to the same database, are scheduled for insertion into the instruction pipeline as two spaced apart instructions. This example assumes the database does not contain the search key. At Blocks 1202 and 1206a, a check is initially performed to determine whether the first instruction is a SEARCH instruction and then a search is made of the I_CACHE to detect the presence of an equivalent search key. For purposes of this example, this search of the I_CACHE results in a negative result, Block 1208a, and control is passed to Block 1224. At Blocks 1224 and 1226, a first SEARCH operation is performed on the CAM core. A MISS result is returned in response to the first SEARCH operation and the I_CACHE is updated with the corresponding search key, Blocks 1228 and 1230. The MISS result in then processed, Block 1232.
Assuming now that the lag time associated the second SEARCH instruction relative to the first SEARCH instruction enables the I_CACHE to be updated before the second SEARCH instruction is inserted into the pipeline, then the second SEARCH instruction results in a search of the I_CACHE, which is performed at Block 1206a. The result of this search indicates the presence of the equivalent search key, Block 1208a. Then, at Block 1210, a check is made to determine whether the duplicate bit associated with the equivalent search key is asserted. Because the duplicate bit has not been set, control is passed to Blocks 1218 and 1220. At Block 1218 the duplicate bit is set and at Block 1220 the second SEARCH instruction is converted into a LEARN instruction. This LEARN instruction is inserted into the instruction pipeline, Block 1224, and then the operations illustrated by Blocks 1226, 1228 and 1232 are performed. At Block 1232, the address of the entry that received the new search key during the LEARN operation is passed to a corresponding results mailbox and the command host is ultimately notified of the entry address corresponding to the second SEARCH instruction. In this manner, the I_CACHE may be used to not only prevent duplicate learn events, as described in the first example, but may also be used in certain circumstances to block repeated MISS results from occurring in response to repeated equivalent search operations. If this feature is not necessary, then the instruction loading and execution logic 524 may be programmed so that the operation illustrated by Block 1202 is not performed and the operations illustrated by Blocks 1206a, 1208a, 1210, 1218 and 1220 are bypassed.
In the drawings and specification, there have been disclosed typical preferred embodiments of the invention and, although specific terms are employed, they are used in a generic and descriptive sense only and not for purposes of limitation, the scope of the invention being set forth in the following claims.
Claims
1. An integrated circuit chip, comprising:
- a CAM-based search engine device that is configured to convert a learn portion of a search-and-learn (SNL) instruction associated with a search key into a search operation using the search key, in response to detecting a prior equivalent learn of the search key in said search engine device.
2. The chip of claim 1, wherein a search portion of the SNL instruction and the search operation result in repeated searches of an internal database using equivalent comparands.
3. The chip of claim 2, wherein a first one of the repeated searches generates a miss result; and wherein a second one of the repeated searches generates a hit result.
4. The chip of claim 1, wherein said search engine device is further configured to prevent a pair of equivalent SNL instructions that result in immediately consecutive searches of an internal database from causing a duplicate learn event.
5. The chip of claim 1, wherein said search engine device comprises a SNL cache memory device that is configured to store search keys that accompany SNL instructions and marker information that identifies whether the stored search keys are duplicates of other search keys within the SNL cache memory device.
6. The chip of claim 5, wherein the SNL cache memory device operates as a first-in first-out (FIFO) memory device.
7. An integrated circuit chip, comprising:
- a CAM-based search engine device that is configured to support conversion of a search-and-learn (SNL) instruction into a pair of equivalent search operations that are pipelined relative to each other when necessary to reduce the occurrence of duplicate learn operations within the search engine device.
8. The chip of claim 7, wherein said search engine device comprises a SNL cache memory device that is configured to store search keys that accompany SNL instructions and marker information that identifies whether the stored search keys are duplicates of other search keys within the SNL cache memory device.
9. The chip of claim 8, wherein the SNL cache memory device operates as a first-in first-out (FIFO) memory device.
10. An integrated circuit chip, comprising:
- a CAM-based search engine device that utilizes a search and learn (SNL) cache memory device as a temporary search key buffer that supports operations to block the occurrence of duplicate learn operations within said search engine device.
11. The chip of claim 10, wherein the SNL cache memory device operates as a first-in first-out (FIFO) memory device.
12. A method of operating a CAM-based search engine device, comprising the steps of:
- searching a database within a CAM core with a search key to detect the presence of a hit or miss condition, in response to a search and learn (SNL) instruction; and
- checking the search key to determine whether is has been marked as a duplicate, in response to detecting the presence of a miss condition.
13. The method of claim 12, wherein said checking step is followed by the step of converting a learn portion of the SNL instruction into a search operation that results in another search of the database with the search key, in response to determining that the search key has been marked as a duplicate.
14. An integrated circuit chip, comprising:
- a CAM-based search engine device that is configured to support internal conversion of a search-and-learn (SNL) instruction into a pair of search operations in order to prevent the SNL instruction from adding a duplicate entry into a database within said search engine device.
15. An integrated circuit chip, comprising:
- a CAM-based search engine device that is configured to support processing of first and second immediately consecutive and equivalent search and learn (SNL) instructions as a first SNL instruction and a second search and search instruction, respectively, in order to block an addition of a duplicate learned entry into a database in said search engine device.
16. The chip of claim 15, wherein said search engine device is further configured to selectively block processing of the second SNL instruction as a second search and search instruction in response to detecting the database as full when the first SNL instruction is processed.
17. The chip of claim 16, wherein said search engine device comprises a cache memory device that operates as a temporary search key buffer that supports operations to process the second SNL instruction as a second search and search instruction in order to block the learning of at least one duplicate entry into the database.
18. The chip of claim 17, wherein the SNL cache memory device operates as a first-in first-out (FIFO) memory device.
19. The chip of claim 17, wherein the SNL cache memory device operates as content addressable memory device that is subject to aging operations.
20. An integrated circuit chip, comprising:
- a CAM-based search engine device that utilizes a searchable cache memory device therein as a search key buffer that supports operations to block duplicate learn events from occurring within said search engine device.
21. The chip of claim 20, wherein the cache memory device comprises a content addressable memory array.
22. The chip of claim 21, wherein the content addressable memory array is configured to store search keys and database identification information.
23. An integrated circuit chip, comprising:
- a CAM-based search engine device having a cache memory device therein, which is configured to retain search keys associated with previously processed learn instructions, and instruction execution logic therein that is configured to process a learn instruction and corresponding search key by searching the cache memory device to detect the presence of a duplicate search key before issuing the learn instruction to a CAM core within said search engine device.
24. The chip of claim 23, wherein the cache memory device comprises a content addressable memory array.
25. A CAM-based search machine, comprising
- a plurality of CAM-based search engine devices that are configured in a depth-cascaded arrangement and include at least one search engine device having a searchable cache memory device therein that is configured to store search keys and database identification information associated with learn instructions previously processed by the search machine and instruction execution logic that is configured to process a learn instruction and corresponding search key by searching the cache memory device to detect the presence of an equivalent search key before issuing the learn instruction to a CAM core in said plurality of CAM-based search engine devices.
26. The search machine of claim 25, wherein the cache memory device comprises a content addressable memory array that is subject to aging.
27. The search machine of claim 25, wherein the at least one search engine device is a master search engine device and the other search engine devices within said plurality of search engine devices are slave search engine devices.
| 6148364 | November 14, 2000 | Srinivasan et al. |
| 6219748 | April 17, 2001 | Srinivasan et al. |
| 6240485 | May 29, 2001 | Srinivasan et al. |
| 20010048687 | December 6, 2001 | Coden |
| 20020085507 | July 4, 2002 | Ku et al. |
| 20020147800 | October 10, 2002 | Gai et al. |
| 20020152209 | October 17, 2002 | Merugu et al. |
| 20030009618 | January 9, 2003 | Regev et al. |
| 20040128455 | July 1, 2004 | Regev |
Type: Grant
Filed: Nov 21, 2003
Date of Patent: Mar 20, 2007
Assignee: Integrated Device Technology, Inc. (San Jose, CA)
Inventors: Jakob Saxtorph (San Jose, CA), John R. Mick, Jr. (Sunnyvale, CA), Harmeet Bhugra (San Jose, CA)
Primary Examiner: Kimberly McLean-Mayo
Attorney: Myers Bigel Sibley & Sajovec PA
Application Number: 10/721,036
International Classification: G06F 12/00 (20060101);
